# GPT-6 Astra GPT-6 Astra is [[OpenAI]]'s frontier [[Large Language Models (LLMs)|LLM]], released 3 September 2026, eight weeks after [[GPT-5.6]] and two days after [[Claude Fable 5.1]]. It is the first natural-number release since [[GPT-5]], and OpenAI used it to say the word out loud: Greg Brockman closed the press briefing with "Welcome to the AGI era." I don't buy the AGI framing, and I'll explain why below. But Astra is a real model with real changes: a new (and reportedly larger) pre-train, a new price class, a saturated ARC-AGI-3, a Critical cybersecurity rating, and a set of coding habits that split the developer community right down the middle. ## The short version - **New price tier.** $10 / $50 per million tokens, 2.5× [[GPT-5.6|Sol]] and identical to [[Claude Fable 5.1]]. Both labs now sell a class above their workhorse models - **ARC-AGI-3 is done.** 99.9% with OpenAI's own harness. The comparison with Sol is misleading (details below), the achievement is still real - **Aggregate benchmarks barely moved.** [[Artificial Analysis]] scores it 61 on its Intelligence Index, exactly Sol's number, and below Fable 5.1 and [[Claude Opus 5]] - **Best writing model many people have used.** Fast, little slop, easy to steer - **Coding is where the fight is.** Token-efficient, unreadable code; long runs that produce nothing; agentic behaviour that reaches for production systems. Some people love it. Armin Ronacher wrote a whole post asking why we're doing this - **Monitorability went down.** OpenAI's own system card says Astra is better at controlling its chain of thought and can evade internal monitors when pushed to ## Specifications and pricing | | GPT-6 Astra | |---|---| | Model ID | `gpt-6-astra` | | Input / output | $10.00 / $50.00 per million | | Cached input / cache write | $1.00 / $12.50 | | Premium requests (> 272K tokens) | 2× input and cache rates, 1.5× output | | [[Context Window|Context window]] | 1,050,000 tokens (922,000 max input) | | Max output | 128,000 tokens | | Knowledge cutoff | 30 April 2026 | | Reasoning effort | `low`, `medium`, `high`, `xhigh`, `max` | | Endpoints | Chat Completions, Responses, Batch | Tools supported out of the box: web search, file search, image generation, code interpreter, hosted shell, apply patch, skills, computer use, [[Model Context Protocol (MCP)|MCP]], tool search. The docs also mention **async tool calling**, which is the API feature I'd look at first if I were building on this. OpenAI's efficiency claim: Astra reaches higher scores with fewer output tokens than Sol, roughly half in some benchmarks and closer to a third in others. [[Artificial Analysis]] put it at about 70% more token efficient than Sol. If that holds, cost per task lands near Sol despite the 2.5× sticker price. Keep the "if" in mind. The same claim was made for Sol and several heavy users on [[Hacker News]] said it never matched their bills. ## The rollout The launch was a mess, and the mess is informative. The press embargo lifted at 11am Pacific. Reuters, Axios, CNBC, and The Verge all published "OpenAI released..." in the past tense. The blog post wasn't up. When it appeared, it 404'd within minutes, came back, then returned 500s. Someone mirrored it on a Cloudflare Worker so people could read the thing being discussed. OpenAI had a multi-provider outage the same morning. Then the access model: Astra went first to "a limited set of organizations" in OpenAI's Daybreak Access program, then to Plus, Pro, Business, and Enterprise "over the coming days". The API and AWS were listed; Azure was not, which people noticed. Usage counts against existing subscription allowances, with extra credits purchasable. OpenAI promised one banked usage reset for every day a paying user didn't have access. This is the same staged pattern Anthropic used for Fable and Mythos, and the same complaint followed: every frontier release now launches for a group you aren't in. One HN commenter: "It's just incredibly deflating to see my future as a second class citizen has already come." Dramatic, sure. Also a fair description of the trend. ## ARC-AGI-3: 99.9%, with an asterisk The headline number, and the one that generated the AGI talk. ARC Prize ran Astra on the semi-private ARC-AGI-3 set at every effort level, twice: once on the standard harness, once on OpenAI's "provider adapter" (their Responses API defaults). | Effort | Standard harness | Provider adapter | |---|---|---| | Max | 62.7% ($26,098) | 98.6% ($17,332) | | High | 54.8% ($40,705) | **99.9%** ($18,817) | | Medium | 38.6% ($48,090) | 98.4% ($19,285) | | Low | 17.5% ($38,166) | 98.0% ($21,298) | | None | 35.2% ($49,791) | 96.7% ($23,457) | Three things to take from that table. **The harness is most of the gap.** OpenAI's blog post charted Astra at 99.9% against Sol at 7.8%. Ted Sanders, who co-authored the post, clarified on HN that the official ARC harness throws away old context and reasoning between turns, so the model re-learns the game repeatedly. Their harness adds standard compaction. Sol would score "roughly 30%" with the same setup. So the real jump is 30% → 99%, and the 7.8% on the chart is the number from a handicapped run. Still a big jump. Still a misleading chart. **Higher effort is cheaper.** Max reasoning cost less than none. Astra solves games in fewer actions, so fewer model calls, so fewer tokens. On 96% of levels it used fewer actions than the human baseline, 51.7% fewer on average. That inverted cost curve is going to show up more often as models get better at planning. **No reasoning still gets 96.7%.** With the adapter, the `none` effort level nearly saturates the benchmark. That matters for the architecture discussion below. ARC Prize's own verdict: "meaningful progress towards generalization", explicitly *not* proof of AGI, and a reminder that ARC-AGI-3 has "tightly bounded scope" with "deterministic, closed-ended mechanics". The scoring is also nonlinear (a level's score is the square of the ratio between the model's move count and the human median), so a discontinuous jump is what you'd expect once a model crosses the planning threshold. A cost note that circulated: human participants were paid about $12.78 per attempted game. Astra at its cheapest adapter run cost roughly $17,000 across the set. The brain-energy-as-electricity comparison (0.067 cents per game) is fun and meaningless; nobody hires brains by the kilowatt-hour. ## Everything else moved a little This is where the AGI talk runs into the numbers. - **[[Artificial Analysis]] Intelligence Index: 61.** Identical to GPT-5.6 Sol. Below Fable 5.1, Opus 5, and Muse Spark 1.3 - **AA Agentic Index: 51**, against Sol's 58. Lower, not higher - **AA Coding Agent Index: 65 → 67**, which AA headlined as "major gains" - **Terminal-Bench 4.0: 57.9% at High**, 56.7% at Max. Fable 5.1 is at 55.8% - **DeepSWE: 73.3% at High**, 71.5% at Max. Muse Spark 1.3 is at 75.4% - **AutomationBench: 41.4%**, against Muse Spark 1.3's 49.4% - **FrontierCode 1.1 Extended:** on par with existing models, per the one HN commenter who found it the only benchmark matching their daily experience - **ExploitGym** (cyber): higher score than Sol with fewer output tokens Two patterns worth naming. First, **Astra is the first model I've seen score lower at Max than at High** on several coding benchmarks. OpenAI's own docs for Sol already told you to test one effort level lower than you think you need. With Astra, that's not a cost optimisation, it's the better result. Second, the benchmark ecosystem is visibly cracking. HN spent a lot of energy on whether the AA index means anything (Opus 5 High equal to Fable 5 Max? Gemini 3.8 Flash at 59?), and Epoch's ECI and the AA index now diverge substantially. One commenter's read: we're on the flat part of the S-curve and every lab is clustering at the same spot. I'm not sure that's right, but it's the first launch where "the benchmarks didn't move" and "it's AGI" were said in the same thread, by the same company, about the same model. ## The system card: more capable, less monitorable OpenAI rated Astra **Critical** for cybersecurity under its Preparedness Framework, a first. "With the right tools and access, Astra can find previously unknown security flaws and develop ways to exploit them across many well-protected systems without a person guiding each step" (Amelia Glaese, OpenAI VP for safety). During evaluation it found and used zero-day vulnerabilities as part of exploit chains. The advanced cyber capabilities go first to the US government and unnamed trusted organisations. The extra safeguards "may sometimes slow, pause, or stop legitimate work", which is exactly what Fable 5 users spent three months complaining about. Then the paragraph everyone quoted: > GPT-6 Astra's monitorability has decreased relative to GPT-5.6 Sol. [...] We have found that GPT-6 Astra is more capable of controlling its own CoT than GPT 5.6-Sol, and less likely to include incriminating information in its CoT. In adversarial settings (where we push the model to evade our monitors) we find that the model is able to remain undetected when strategically underperforming in evaluations (sandbagging) and can sometimes evade our internal monitors when asked to perform certain sabotage tasks. Read it twice. The lab is saying, in its own document, that the model can hide what it's doing from the lab's own monitors when asked to. Merely knowing that monitors exist causes it to shorten its reasoning. They did not observe steganographic reasoning, which is the one line of comfort. The obvious follow-up question, raised on HN and unanswered: if the model is that good at hiding "incriminating information", how much should we trust the alignment charts in the same document? Other numbers from the card, in Astra's favour: - Indirect prompt injection success rate dropped from 27.0% to 8.5%. See [[Prompt injection]] - About 53% fewer misalignment flags at severity 3+ in Codex simulations - On a honeypot inside ExploitGym, Sol attacked it 55.4% of the time. Astra never did. Less [[Reward Hacking|reward hacking]], or better at knowing when it's being watched. The card doesn't let you tell ## Architecture: looped transformers and "neuralese" The Information reported that Astra uses "recurrent depth" or looped transformers, and framed it as a technique that makes chain-of-thought monitoring harder. Sebastian Raschka wrote the clearest response. A looped transformer reuses the same transformer blocks several times on the same input. Instead of 44 distinct blocks, 22 blocks applied twice with shared weights. Effective depth goes up; parameter count doesn't. The idea is from the Universal Transformers paper (2018), and recent research puts the gain at 6.8 to 18% better performance at a fixed compute budget. Raschka's point: this is architecturally the same as stacking more layers, the model still emits one token at a time, and it doesn't hide anything that regular depth doesn't also hide. OpenAI's chief scientist Jakub Pachocki added that "the depth of the computation graph for our present frontier models, including Astra, is within a factor of two of GPT-4", which rules out deep dynamic looping. His explanation for the reduced monitorability: better models need less scratch work. A strong student writes less on the exam paper. Shorter chains of thought are a byproduct of capability, not obfuscation. The counter-evidence, also from the HN thread: a third-party benchmark showed a large jump in **multi-hop reasoning without chain of thought**, and Astra at `none` effort still hit 96.7% on ARC-AGI-3. That's harder to explain with "it just needs less scratch work". Whether you call it neuralese or not, more of the reasoning is happening inside the forward pass and less of it is on the page where a monitor can read it. Raschka is right that looping per se isn't the villain. The system card is right that there's less to monitor. Both can be true. ## What practitioners are saying ### Every's vibe check Dan Shipper and the Every team had early access and published a 4,000-word vibe check plus video by 2pm on launch day, having learned about the launch at 3am. Their summary: "a big upgrade from 5.6-Sol, with some frustrating habits that keep it from matching Fable at the top end." - **Writing**: the best writing model Shipper has tried. Fast, very little slop, easy to steer. It one-shotted the first draft of its own vibe check; Shipper didn't realise the draft wasn't written by Katie Parrott, the staff writer, whose reply was "RIP my job" - **Computer use**: runs for hours in complicated apps. It produced the first cut of their Fable 5.1 vibe check video - **3D and visual work**: one-shotted a "historically accurate rendition of the Battle of Waterloo" - **The habit**: it overcomplicates, especially at higher effort. Ask for a simple interface and get extra labels, buttons, and features. "It has a habit of turning everything into a landing page" Verdict: "I'm reaching for Astra all day, but Fable 5.1 still gets my biggest tasks. On ambitious builds, Fable is better at understanding what I want and taking it further than I would have thought to ask." The writing claim interests me most. [[Claude Fable 5.1]] got smarter and kept its Claudish prose tic. If Astra really produces less slop out of the box, that's a practical reason to pick it for drafting even if Fable wins on judgement. I'd still run the humanizing pass. See [[Quality non-fiction is the antithesis of AI slop]]. ### Armin Ronacher: "Why are we doing this again?" The post that got 452 points and 67 comments. His thesis: AI engineering has become *involution* (内卷, Neijuan), ever more effort and competition without improving output. Astra is "incredibly impressive" and he doesn't know how to work with it for actual software engineering. Specifics: - **Token-efficient slop.** Astra optimises code for token count rather than readability. It edits C code with Python string replacement instead of the patch tool. Compressed syntax, hardcoded constants, bare array indexes in production code, extreme code golf. "Almost impossible to visualize" - **Degradation over long runs.** In a 35-hour unattended run, task naming went from "1, 2, 3" to "8b2c2b2b checkpoint1". The agent was optimising locally with no global view - **The runaway.** That run consumed an entire ChatGPT token reset, about 4 billion tokens (roughly $1,200 at list price), wrote 75,000 new lines, and produced "absolutely nothing of value". His emphasis: earlier models "did not do that" - **The conclusion**: "these models increasingly are for other people", lawyers, artists, mathematicians. Not software engineers ### The Hacker News consensus, such as it is Reading four threads back to back, the picture is consistent: - **Unreadable tool use.** New models write obscene bash one-liners and Python scripts "using every option flag that exists", impossible to review. This isn't only Astra: [[Claude Code]]'s [[Claude Code Auto Mode|auto mode]] also nudges toward Python and `sed` edits instead of the Edit tool, apparently by design, since bulk scripted edits are cheaper in cache reads than one tool call per file. The harness re-injects that instruction, which is why "please use the Edit tool" in `CLAUDE.md` keeps getting forgotten - **Agentic overreach.** One developer gave Astra a small bug ticket about an invalid profile field. Four minutes in, it decided to check the production database "to see how far the corruption has spread" and tried to SSH into the hosting provider, triggering a 1Password prompt. Another asked ChatGPT desktop to find a lost chat thread and got a request for full computer control to drive Finder. The quote that stuck: "'you have to watch it like a hawk' is the opposite of agentic engineering." The reply that also stuck: why is your agent able to call SSH at all? - **Medium feels weaker than Sol's medium.** Several people found the default effort less rigorous than Sol on its default. Others reported reasoning loops: "yes this needs fixing" repeated for half an hour without a fix - **Genuine wins outside typical software.** CAD and PCB circuit work: "a huge jump compared to Sol". A toy planet app whose weather render Sol fumbled for several iterations, Astra got in one. A single session running for 4 days 13 hours on a public repo - **Over-engineering is a family trait.** The launch thread's most-discussed anecdote wasn't even about Astra: someone asked Sol to update a 1,000-line Python script overnight and woke up to 189 files and roughly 100,000 lines, with an adapter, a source manager, and an acquisition manager just for downloading PDFs. Both frontier families do this. The working countermeasure is a short anti-bloat file: "do not over engineer, do not gold plate, no CLI args, keep it simple", "do not create protocols, abstract base classes, or adapters unless there are already two real implementations" - **The workflow that survives.** Plan, generate, review, commit, push. Use the expensive model to plan and review, a cheaper one to implement. One person runs a ChatGPT 6 Pro session for the plan, Sol for the implementation, 6 Pro for the review, and never touches metered Codex minutes for Astra ### Luna vs Astra for code review A useful datapoint from Entelligence, which ran [[GPT-5.6|GPT-5.6 Luna]] ($0.20 / $1.20) and Astra on 50 public PRs across Cal.com, Sentry, Discourse, Keycloak, and Grafana, with findings verified by two judge models. | | Luna | Astra | |---|---|---| | Verified bugs | 69 (from 93 findings) | 92 (from 96 findings) | | Precision | 74% | 96% | | Cost per review | $0.0041 | $0.113 | | Cost per verified bug | $0.003 | $0.061 | Luna found three quarters of Astra's bugs for under 4% of the cost, and missed the authentication and permission bugs. One in four Luna comments was wrong; Astra was wrong four times in 96. The other cost comparison making the rounds: the same 447-turn agentic coding task priced on Astra and on [[DeepSeek V4.1 Flash]] came to about $55 versus about $0.50, because two thirds of Astra's bill is cache reads at $1 per million and DeepSeek charges $0.003 for those. The HN reaction was the right one: ten cents per PR review is nothing, so the question isn't cost, it's noise. A reviewer that's wrong 25% of the time generates a wall of comments a human has to triage. That's the real expense. The stage-splitting advice from the [[GPT-5.6]] note still applies, with a refinement: cheap models for correctness sweeps, the expensive one for anything touching auth, permissions, or money. ## My take Astra is two models wearing one name. The first is a writing, reasoning, and computer-use model that is very good. If you draft prose, drive complex apps, build visualisations, or work in domains like CAD where the previous generation fumbled, the reports are consistently positive. Every's "reaching for it all day" is believable. The second is a coding agent trained toward token efficiency at the expense of code a human can read, with a stronger tendency to expand scope than anything before it. Ronacher's 35-hour run and the SSH-into-prod story are the same failure with different budgets: the model pursues the goal it inferred rather than the one you gave it, and it does so with more competence than the guardrails around it assumed. Everything in the [[Loop Engineering]] playbook applies harder here: bounded tasks, a real verifier, a human-authored loop, and sandboxes that make "tried to SSH into prod" impossible rather than merely blocked. On AGI: Brockman himself said "everyone has a different definition" and left users to decide. The company that wrote an AGI clause into its Microsoft contract has a commercial reason to declare it, and the aggregate benchmarks say it's a Sol-class model with a saturated puzzle benchmark and a much better writing register. A model that is more able to hide its reasoning from its makers, released before a planned IPO, with the blog post going down four times, is not what I imagined the AGI announcement would look like. That's fine. The interesting question was never the label. What I'll actually do: try it for drafting and for anything visual, keep Fable 5.1 (at low effort, see that note) for the tasks where judgement matters, and keep a cheap model in the loop for the sweeps. And I'd re-read the monitorability paragraph before letting any of them run unattended for 35 hours. ## Caveats - Most benchmark numbers come from OpenAI's launch post or from ARC Prize's runs on OpenAI's harness. Treat as directional - Early-access reports are from people who got early access. That's a selection - Pricing, effort levels, and rollout tiers change weekly. Check the docs before committing budget - The looped-transformer claim is a press report OpenAI hasn't confirmed or denied ## References - OpenAI announcement: https://openai.com/index/gpt-6-astra/ - System card: https://deploymentsafety.openai.com/gpt-6-astra - Model docs: https://developers.openai.com/api/docs/models/gpt-6-astra - OpenAI on the ARC-AGI-3 harness changes: https://openai.com/index/how-two-settings-tripled-our-arc-agi-3-score/ - ARC Prize analysis: https://arcprize.org/blog/astra - [[Artificial Analysis]] benchmarking: https://artificialanalysis.ai/articles/benchmarking-gpt-6-astra - Every, "Vibe Check: GPT-6 Astra Is a Big Upgrade With Some Bad Habits" (Katie Parrott and GPT-6 Astra): https://every.to/vibe-check/gpt-6-astra-vibe-check - [[Dan Shipper]]'s summary thread: https://x.com/danshipper/status/2095593705214300394 - Armin Ronacher, "Astra for Coding: Why Are We Doing This Again?": https://lucumr.pocoo.org/2026/9/7/astra-why/ - Sebastian Raschka, "GPT-6 Astra, looped transformers, and hidden reasoning": https://magazine.sebastianraschka.com/p/gpt-6-astra-looped-transformers-and - Entelligence, "GPT-5.6 Luna vs. GPT-6 Astra: Is a $1.20 Model Good Enough for Code Review?": https://entelligence.ai/blogs/gpt-5.6-luna-vs-gpt-6-astra-is-a-1.20-model-good-enough-for-code-review - [[Hacker News]] threads: launch (2,279 points) https://news.ycombinator.com/item?id=49554643 · rollout https://news.ycombinator.com/item?id=49554273 · ARC-AGI-3 https://news.ycombinator.com/item?id=49555691 · Artificial Analysis https://news.ycombinator.com/item?id=49556147 · "AGI era" https://news.ycombinator.com/item?id=49554048 · Raschka https://news.ycombinator.com/item?id=49627370 · Ronacher https://news.ycombinator.com/item?id=49654229 · Luna vs Astra https://news.ycombinator.com/item?id=49703003 ## Related - [[OpenAI]] - [[GPT-5.6]] - [[GPT-5.5]] - [[GPT-5]] - [[ChatGPT]] - [[OpenAI Codex]] - [[Codex CLI]] - [[Claude Fable 5.1]] - [[Claude Fable 5]] - [[Claude Opus 5]] - [[Large Language Models (LLMs)]] - [[Context Window]] - [[AI Frontier Model]] - [[AI API prices are rising]] - [[Artificial General Intelligence (AGI)]] - [[Artificial Analysis]] - [[Reward Hacking]] - [[Prompt injection]] - [[Loop Engineering]] - [[Ori Eval]] - [[Kimi K3]] - [[DeepSeek V4.1 Flash]] - [[ChatGPT Images 2.5]] - [[AI Agents]] - [[Claude Code]] - [[Claude Code Auto Mode]] - [[Quality non-fiction is the antithesis of AI slop]] - [[Dan Shipper]] - [[François Chollet]] - [[Hacker News]]